RESEARCH ON PHYSICAL FITNESS

TRAINING OF FOOTBALL PLAYERS BASED

ON IMPROVED LSTM NEURAL NETWORK

TO IMPROVE PHYSICAL ENERGY SAVING

AND HEALTH

Nengchao Pan*

Department of Basic Studies, Beihai Vocational College, Beihai, Guangxi, 536000,

China

bhzypnc@163.com

Reception: 05/11/2022 Acceptance: 06/01/2023 Publication: 02/02/2023

Suggested citation:

P., Nengchao. (2023). Research on physical ﬁtness training of football

players based on improved LSTM neural network to improve physical

energy saving and health. 3C Tecnología. Glosas de innovación aplicada a

la pyme, 12(1), 127-140. https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

127

ABSTRACT

In order to ensure that the physical function of football players adapts to the

development of modern football level, and avoid the phenomenon of inability to adapt

to the intensity of modern football games due to lack of physical fitness. Aiming at the

physical training of football players, this paper proposes an improved long-short-term

memory network (W-LSTM) model for the optimization and prediction of physical

training. The model effectively combines the global feature extraction ability of LSTM

for time series data and the preprocessing ability of the extracted data, which reduces

the loss of feature information and achieves high prediction accuracy. The front door

is added on the basis of LSTM, which combines training and physical function to

reduce the impact of fluctuations in data outliers on the prediction results, effectively

improving the accuracy of physical training optimization and prediction, and using

body shape, exercise tolerance, exercise intensity and fitness level as input values to

conduct comparative experiments on the three models of W-LSTM, LM-BP and

ARIMA. The study found that W-LSTM has a lower mean square error (0.011) and a

higher correlation coefficient (0.985), indicating that the model proposed in this paper

is significantly better than other existing comparison models in terms of the accuracy

of prediction results.

KEYWORDS

W-LSTM; Football; Athlete; Physical fitness; Prediction model

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

128

ABSTRACT

In order to ensure that the physical function of football players adapts to the

development of modern football level, and avoid the phenomenon of inability to adapt

to the intensity of modern football games due to lack of physical fitness. Aiming at the

physical training of football players, this paper proposes an improved long-short-term

memory network (W-LSTM) model for the optimization and prediction of physical

training. The model effectively combines the global feature extraction ability of LSTM

for time series data and the preprocessing ability of the extracted data, which reduces

the loss of feature information and achieves high prediction accuracy. The front door

is added on the basis of LSTM, which combines training and physical function to

reduce the impact of fluctuations in data outliers on the prediction results, effectively

improving the accuracy of physical training optimization and prediction, and using

body shape, exercise tolerance, exercise intensity and fitness level as input values to

conduct comparative experiments on the three models of W-LSTM, LM-BP and

ARIMA. The study found that W-LSTM has a lower mean square error (0.011) and a

higher correlation coefficient (0.985), indicating that the model proposed in this paper

is significantly better than other existing comparison models in terms of the accuracy

of prediction results.

KEYWORDS

W-LSTM; Football; Athlete; Physical fitness; Prediction model

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

PAPER INDEX

ABSTRACT

KEYWORDS

1. INTRODUCTION

2. W-LSTM RELATED MODEL THEORY

2.1. Data preprocessing

2.2. Definition of the model

2.3. Training process

3. TEST AND RESULT ANALYSIS

3.1. Experiment Setup Instructions

3.2. Simulation comparison test

4. CONCLUSION

5. CONFLICT OF INTEREST

REFERENCES

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

129

1. INTRODUCTION

With the increasing level of football around the world, the competition is

intensifying, and the level of scientific training is also increasing. The daily training

experience of football players around the world shows that athletes in different periods

and regions follow some almost the same training principles in terms of training

content and training methods, and have some similar or even common characteristics

[1]. However, with the diversification of the world football level and training process,

this lack of training content and training methods for individual characteristics has

been unable to adapt to the inevitable laws of football development, and will inevitably

hinder the improvement of athletes' competitive ability and regional football level [2].

Therefore, while we carry out the overall general training of football players, it is

absolutely necessary to carry out individualized training for players, and it is also in

line with the general law of the development of world football.

The physical training of football players is a time series problem for improving

physical energy and health. For the prediction problem of time series, long-short term

memory neural network [3] (long-short term memory, LSTM) has been widely used in

speech recognition [4], network Flow prediction [5], pre-drilling logging curve

prediction[6], power and image prediction [7-8], toxic gas law prediction[9] and other

fields. Mao et al. proposed an LSTM model for image caption generation as early as

2015, pioneering the application of this research field in image caption generation.

Peng et al. [10] used LSTM for the prediction of generated sentences, using dual

LSTM layers to tune the parameters to improve the accuracy of sentence generation.

In 2017, some scholars proposed a new time-varying parallel recurrent neural network

for the generation of sports health image captions, which can obtain dynamic visual

and textual representations at each time step, thus solving existing methods. The

problem that currently generated words do not match the obtained image features in

[11]. In addition, some scholars have applied the attention mechanism to the

prediction of physical education innovation indicators, and found that the attention

model can effectively improve the prediction accuracy of the innovation direction of

physical education [12]. Kyunghyun et al. [13-14] proposed another gating mechanism

of Gated Recurrent Unit (GRU), which is different from LSTM. The goal is to make

each recurrent unit adaptively capture the dependencies of different time scales.

Chung et al. [15] also conducted a specific study on GRU. However, this idea is also

difficult to process data in combination with abnormal fluctuations and large

fluctuations of data.

Physical fitness is one of the five basic elements of football players' competitive

ability, and it is the physical ability necessary for football players to perform their

technical and tactical skills normally and achieve excellent sports performance [16].

Physical fitness plays a pivotal role in a competitive football game. However, each

athlete's upper limit of physical fitness and reserves are not the same, so it is difficult

to excavate the limit of each athlete if the traditional unified training method is used

[17]. In this regard, this study addresses the importance of physical fitness training

using the LSTM model. However, the traditional long-term memory neural network

model has the problem of premature saturation. Therefore, considering the

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

130

1. INTRODUCTION

With the increasing level of football around the world, the competition is

intensifying, and the level of scientific training is also increasing. The daily training

experience of football players around the world shows that athletes in different periods

and regions follow some almost the same training principles in terms of training

content and training methods, and have some similar or even common characteristics

[1]. However, with the diversification of the world football level and training process,

this lack of training content and training methods for individual characteristics has

been unable to adapt to the inevitable laws of football development, and will inevitably

hinder the improvement of athletes' competitive ability and regional football level [2].

Therefore, while we carry out the overall general training of football players, it is

absolutely necessary to carry out individualized training for players, and it is also in

line with the general law of the development of world football.

The physical training of football players is a time series problem for improving

physical energy and health. For the prediction problem of time series, long-short term

memory neural network [3] (long-short term memory, LSTM) has been widely used in

speech recognition [4], network Flow prediction [5], pre-drilling logging curve

prediction[6], power and image prediction [7-8], toxic gas law prediction[9] and other

fields. Mao et al. proposed an LSTM model for image caption generation as early as

2015, pioneering the application of this research field in image caption generation.

Peng et al. [10] used LSTM for the prediction of generated sentences, using dual

LSTM layers to tune the parameters to improve the accuracy of sentence generation.

In 2017, some scholars proposed a new time-varying parallel recurrent neural network

for the generation of sports health image captions, which can obtain dynamic visual

and textual representations at each time step, thus solving existing methods. The

problem that currently generated words do not match the obtained image features in

[11]. In addition, some scholars have applied the attention mechanism to the

prediction of physical education innovation indicators, and found that the attention

model can effectively improve the prediction accuracy of the innovation direction of

physical education [12]. Kyunghyun et al. [13-14] proposed another gating mechanism

of Gated Recurrent Unit (GRU), which is different from LSTM. The goal is to make

each recurrent unit adaptively capture the dependencies of different time scales.

Chung et al. [15] also conducted a specific study on GRU. However, this idea is also

difficult to process data in combination with abnormal fluctuations and large

fluctuations of data.

Physical fitness is one of the five basic elements of football players' competitive

ability, and it is the physical ability necessary for football players to perform their

technical and tactical skills normally and achieve excellent sports performance [16].

Physical fitness plays a pivotal role in a competitive football game. However, each

athlete's upper limit of physical fitness and reserves are not the same, so it is difficult

to excavate the limit of each athlete if the traditional unified training method is used

[17]. In this regard, this study addresses the importance of physical fitness training

using the LSTM model. However, the traditional long-term memory neural network

model has the problem of premature saturation. Therefore, considering the

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

improvement of the standard LSTM model, a new W-LSTM model is proposed to input

physical fitness data and physical function data to train the model to reduce the

number of different athletes. The influence of physical fitness and physical fitness on

the prediction results, so as to provide a suitable numerical basis for the training of

different athletes.

2. W-LSTM RELATED MODEL THEORY

2.1. DATA PREPROCESSING

Since the athlete's physique is affected by training and physical function, the

following definitions are made: define t as the number of training days: xi represents

the physical fitness data of the i-th day, then the input data is

，

, and so on, until the

last day; the physical energy fluctuation is calculated using the formula

, when i = 1, assuming that the data of the previous day is 0,

then Δ

w1 = w1, the input format for physical fitness is

，

, and

so on until the last day.

2.2. DEFINITION OF THE MODEL

The input to the LSTM model consists of trained and physical performance data,

i.e. using the data from the previous t days as input to predict physical performance

on day t + 1 [18]. LSTM is a special RNN structure, which was proposed by Hochreiter

et al. [19] in 1997 to decide when and how to update the hidden state of RNN. Due to

its unique design structure, LSTM can solve the gradient very well. Disappearance

problem, it is especially suitable for dealing with timing problems. Standard LSTM

units include forget, input, and output gates [20]. On the basis of LSTM, W-LSTM

processes its input information accordingly, and takes training and physical function

data as data input, and it also includes pre-gate, forget gate and output gate [21] (Fig.

1), therefore, it can process more information than a standard LSTM, and its input in

this study contains training and physical performance information.

Xi={x1，x2……xt}

X2={x2，x3……xt+1}

Δpi=pi−pi−1

ΔPi={Δp1，Δp2……Δpt}

ΔP2={Δp2，Δp3，……Δpt+1}

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

131

Figure 1. W-LSTM structure

Front gate, which combines body function information and function change

information to form combined information:

Among them, Xi

is the body function used to analyze the law of physical changes,

Pi is the function fluctuation information extracted from the physical energy information

as input alone to strengthen the model's processing of physical energy fluctuations,

are network parameters [22]. The output result of the tanh

activation function is between [-1, 1]. The closer the output value is to -1, the greater

the negative fluctuation; the closer the output value is to 1, the greater the positive

fluctuation. The larger the data fluctuation, the greater the impact on the training of the

W-LSTM model. On the contrary, when the fluctuation is 0, the input fluctuation data

has no effect on the training of the model. At this time, the W-LSTM model is

equivalent to the standard LSTM [23].

Forget gate is the historical state information that controls whether to “forget” [24].

Among them, ht - 1 is the hidden state of the previous sequence, and Sk

is the input

sequence of this time. Define Wf as the weighted matrix of ht - 1, Uf

as the weighted

matrix of Sk, and bf as the bias.

The Input gate is responsible for supplementing the current input to the latest

"memory". It consists of two parts: first, the Sigmoid layer outputs it; second, a Tanh

layer creates a new candidate value vector, which will be added into the state. Define

{Wt,Ut,bt}{Wa,Ua,ba} as the network parameters of the input gate, then

Then update the cell state:

C

t-1

h

t-1

σtanhσ

i

t

tanh

σ

O

t

h

t

C

t

C*

C

t

h

t

tanh

S

k

X

k

P

k

Front door

(1)

xk i px i b tanhs P

W X P×+ + ×=（

W+b）

{Wx,Wp,bx,bp}

(2)

ƒ

ƒ 1

ƒ

(w U b )

t f k f

h sσ

−

=×+×+

(3)

(4)

1

(w U b )

t t t i i t

i h sσ

−

=×+×+

1

*

a a a

(nh wta U b )

t k

hC s

−

×+×+=

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

132

Figure 1. W-LSTM structure

Front gate, which combines body function information and function change

information to form combined information:

Among them, Xi is the body function used to analyze the law of physical changes,

Pi is the function fluctuation information extracted from the physical energy information

as input alone to strengthen the model's processing of physical energy fluctuations,

are network parameters [22]. The output result of the tanh

activation function is between [-1, 1]. The closer the output value is to -1, the greater

the negative fluctuation; the closer the output value is to 1, the greater the positive

fluctuation. The larger the data fluctuation, the greater the impact on the training of the

W-LSTM model. On the contrary, when the fluctuation is 0, the input fluctuation data

has no effect on the training of the model. At this time, the W-LSTM model is

equivalent to the standard LSTM [23].

Forget gate is the historical state information that controls whether to “forget” [24].

Among them, ht - 1 is the hidden state of the previous sequence, and Sk is the input

sequence of this time. Define Wf as the weighted matrix of ht - 1, Uf as the weighted

matrix of Sk, and bf as the bias.

The Input gate is responsible for supplementing the current input to the latest

"memory". It consists of two parts: first, the Sigmoid layer outputs it; second, a Tanh

layer creates a new candidate value vector, which will be added into the state. Define

{Wt,Ut,bt}{Wa,Ua,ba} as the network parameters of the input gate, then

Then update the cell state:

C

t-1

h

t-1

σtanhσ

i

t

tanh

σ

O

t

h

t

C

t

C*

C

t

h

t

tanh

S

k

X

k

P

k

Front door

(1)

xk i px i b tanhs P

W X P×+ + ×=（

W+b）

{Wx,Wp,bx,bp}

(2)

ƒ

ƒ 1

ƒ

(w U b )

t f k f

h sσ

−

=×+×+

(3)

(4)

1

(w U b )

t t t i i t

i h sσ

−

=×+×+

1

*

a a a

(nh wta U b )

t k

hC s

−

×+×+=

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

The output gate controls how much "memory" can be used in the update of the

next layer of the network. Define

as the network parameters of the

output gate, and the calculation of the output gate can be expressed by formula 6:

After calculating Ot

, it is necessary to use the Tanh function to suppress the

memory value to [-1, 1], so the output formula of the final output gate is:

The historical information output by the last W-LSTM layer passes through a

prediction layer and outputs the result y:

2.3. TRAINING PROCESS

The training process of W-LSTM is as follows: Calculate the output value of the W-

LSTM cell according to the forward calculation formulas (1) ~ (8) [25]; Backpropagate

in two directions according to time and network level to calculate the error term;

according to the corresponding, calculate the gradient of each weight, and update the

weight; repeat (1) to (3) to obtain a set of optimal parameters and keep them. To

prevent overfitting during training, this study uses the Dropout regularization technique

[26], which was proposed by Prof. Hinton's team in 2014. Dropout provides a clever

way to increase the generalization ability of a network model by reducing weight

connections.

3. TEST AND RESULT ANALYSIS

3.1. EXPERIMENT SETUP INSTRUCTIONS

In this section, the proposed W-LSTM model will be evaluated experimentally. The

experimental environment is: INTEL Corei5 CPU, 2.80GHz; 4G memory. The

experimental data is the daily training data of a football team in Xi'an from April 2022

to May 2022. Each comparative experiment was run 10 times, and the average value

was taken.

Three comparison models are set up:

(1) W-LSTM model, input historical function information and physical fitness

fluctuation information to train the model to make predictions.

(5)

t -

*

t 1 ƒt i

iC CC ×+×=

{Wo,Uo,bo}

(6)

1

(w U b )

t o k o t o

O s hσ

−

=×+×+

(7)

t

tanh(C )

t t

h O=×

(8)

t

y W h b=×+

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

133

(2) The BP neural network improved by the LM algorithm only takes the historical

functional information as input, and uses the physical fitness information of the

previous n days to predict the physical fitness situation of the n+1th day.

(3) The ARIMA model regards the data sequence generated by physical fitness

over time as a random sequence, and uses a certain mathematical model to

approximately describe this sequence.

At the same time, in order to test the universality of the W-LSTM model, the three

models were compared using four data of body shape, exercise tolerance, exercise

intensity and fitness level.

3.2. SIMULATION COMPARISON TEST

This summary uses the W-LSTM model, the BP neural network improved by the

LM algorithm and the ARIMA model to conduct experiments, and the mean square

error (MSE) and the coefficient of determination (R2) are used to determine the

accuracy of the prediction results.

MSE and R2 are commonly used indicators to evaluate the accuracy of the model.

MSE is a measure that reflects the degree of difference between the estimator and

the estimated value. The smaller the MSE, the higher the accuracy of the model; the

larger the R2, the greater the difference between the independent variable and the

dependent variable. The higher the degree of explanation, the higher the percentage

of changes caused by independent variables in the total change, and the denser the

observation points are near the regression line, which means the higher the model fit.

Where n represents the total sample, Y_actual represents the real data, Y_predict

represents the prediction result, and Y_mean represents the average value of the real

data.

(9)

(10)

2

2

2

2

2

:

(Y_actual _predict)

:

(Y_actual _predict)

1(Y_actual _mean)

MSE

Y

MSE

n

R

Y

R

Y

−

=

−

=−−

∑

∑

∑

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

134

(2) The BP neural network improved by the LM algorithm only takes the historical

functional information as input, and uses the physical fitness information of the

previous n days to predict the physical fitness situation of the n+1th day.

(3) The ARIMA model regards the data sequence generated by physical fitness

over time as a random sequence, and uses a certain mathematical model to

approximately describe this sequence.

At the same time, in order to test the universality of the W-LSTM model, the three

models were compared using four data of body shape, exercise tolerance, exercise

intensity and fitness level.

3.2. SIMULATION COMPARISON TEST

This summary uses the W-LSTM model, the BP neural network improved by the

LM algorithm and the ARIMA model to conduct experiments, and the mean square

error (MSE) and the coefficient of determination (R2) are used to determine the

accuracy of the prediction results.

MSE and R2 are commonly used indicators to evaluate the accuracy of the model.

MSE is a measure that reflects the degree of difference between the estimator and

the estimated value. The smaller the MSE, the higher the accuracy of the model; the

larger the R2, the greater the difference between the independent variable and the

dependent variable. The higher the degree of explanation, the higher the percentage

of changes caused by independent variables in the total change, and the denser the

observation points are near the regression line, which means the higher the model fit.

Where n represents the total sample, Y_actual represents the real data, Y_predict

represents the prediction result, and Y_mean represents the average value of the real

data.

(9)

(10)

2

2

2

2

2

:

(Y_actual _predict)

:

(Y_actual _predict)

1(Y_actual _mean)

MSE

Y

MSE

n

R

Y

R

Y

−

=

−

=−−

∑

∑

∑

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

Figure 2. W-LSTM prediction results

By modeling and predicting the physical function sequence, Figures 2 and 3 are the

comparison between the prediction results obtained by the W-LSTM model and the

ARIMA model and the actual data. Obviously, for the W-LSTM model, the

experimental value and the predicted value are extremely coincident and very close,

which shows that the W-LSTM model proposed in this study has better prediction

results.

By observing the data in Figure 3, it is found that the experimental results of the

ARIMA model deviate significantly from other models, the coincidence rate between

the experimental values and the predicted values is low, and R2 is even less than 0,

which means that the predicted results have nothing to do with the original data. The

ARIMA model performs well when dealing with stationary time series. When the data

is not stationary, a stationary sequence needs to be obtained through a certain

processing method. The physical function data used in this experiment has

continuous invariance and mutation, that is, continuous invariance within a period of

time. Change, the initial stage gradually increases, this characteristic leads to the loss

of too much information when the data is differentiated, resulting in an extremely poor

prediction effect of the ARIMA model and a large deviation.

0 10 20 30 40 50

1

2

3

4

5

6

7

Functional fitness coefficient

t (d)

Measured

predict

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

135

Figure 3. ARIMA model prediction results

After the training, the input data is used for prediction, and the MSE of the W-LSTM

model and the BP neural network improved by the LM algorithm changes as the

prediction progresses, as shown in Figure 4. The MSE of the W-LSTM model is 0.032,

the MSE of the LM-BP model is 0.059, and the MSE of the ARIMA model is 0.923.

The MSE of the W-LSTM model is smaller than other models, and the model has the

highest accuracy; while R is larger than other models, which means that the fitting

degree of the W-LSTM model is higher than that of other models. In general, the MSE

trends of the two models are roughly the same, and the MSE of the W-LSTM model is

generally smaller than the MSE of the BP neural network improved by the LM

algorithm [27].

Figure 4. W-LSTM model and BP neural network improved by LM algorithm

Changes in MSE

0 10 20 30 40 50

1

2

3

4

5

6

7

Functional fitness coefficient

t (d)

Measured

predict

0 10 20 30 40 50

0.00

0.02

0.04

0.06

0.08

0.10

0.12

0.14

0.16

0.18

0.20

Functional fitness coefficient

t (d)

Measured

predict

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

136

Figure 3. ARIMA model prediction results

After the training, the input data is used for prediction, and the MSE of the W-LSTM

model and the BP neural network improved by the LM algorithm changes as the

prediction progresses, as shown in Figure 4. The MSE of the W-LSTM model is 0.032,

the MSE of the LM-BP model is 0.059, and the MSE of the ARIMA model is 0.923.

The MSE of the W-LSTM model is smaller than other models, and the model has the

highest accuracy; while R is larger than other models, which means that the fitting

degree of the W-LSTM model is higher than that of other models. In general, the MSE

trends of the two models are roughly the same, and the MSE of the W-LSTM model is

generally smaller than the MSE of the BP neural network improved by the LM

algorithm [27].

Figure 4. W-LSTM model and BP neural network improved by LM algorithm

Changes in MSE

0 10 20 30 40 50

1

2

3

4

5

6

7

Functional fitness coefficient

t (d)

Measured

predict

0 10 20 30 40 50

0.00

0.02

0.04

0.06

0.08

0.10

0.12

0.14

0.16

0.18

0.20

Functional fitness coefficient

t (d)

Measured

predict

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

The input data fluctuates greatly at one-third of the total data, and at this time, the

MSE of both models has a short-term increase, while the MSE fluctuation of the W-

LSTM model is smaller than that of the BP neural network improved by the LM

algorithm. This shows that the W-LSTM model has a better effect on handling large

fluctuations in data. In order to verify the universality of the W-LSTM model, four kinds

of data of body shape, exercise tolerance, exercise intensity and health level are used

as input to carry out comparative experiments on the three models. The experimental

results are shown in Table 1. The table shows the evaluation index of the prediction

results of the three models on the four factors respectively. It can be seen that W-

LSTM has better results than other models, and it shows that the W-LSTM model has

good universality [28-29].

Table 1. MSE and R2comparison

It can be seen from the above experiments that the W-LSTM model has higher

accuracy, better fitting degree and good universality. On the whole, W-LSTM is a good

prediction model for the physical fitness prediction problem of football players.

4. CONCLUSION

This paper proposes a physical fitness prediction method for football players based

on the W-LSTM network model. The model is mainly constructed by LSTM, which can

effectively extract the local and global features of influencing factors. After data

analysis and variable reordering based on the maximum information coefficient

method, making the data distribution more regular and easy to train. The research

compares the prediction results of W-LSTM, LM-BP, and ARIMA models. The

experimental results show that: (1) the prediction accuracy of W-LSTM is significantly

better than other methods LM-BP and ARIMA models; (2) W-LSTM has lower MSE

and higher R2 compared to the other two models, the correlation coefficient of its

body shape reaches 0.985; (3) The LSTM is improved to become W-LSTM, and it is

of practical value to apply it to the physical fitness prediction of football players.

MSE

Model Body shape Exercise tolerance Exercise intensity Fitness level

W-LSTM 0.074 0.031 0.137 0.011

LM-BP 0.163 0.048 0.238 0.017

ARIMA 3.244 -0.891 4.618 0.9048

R2

W-LSTM 0.985 0.867 0.904 0.935

LM-BP 0.967 0.793 0.833 0.919

ARIMA 0.317 -2.442 -2.191 -2.771

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

137

5. CONFLICT OF INTEREST

The authors declared that there is no conflict of interest.

REFERENCES

(1) Bangsbo, J., Mohr, M., & Krustrup, P. (2006). Physical and metabolic

demands of training and match-play in the elite football player. J Sports Sci,

24(7), 665-674. https://doi.org/10.1080/02640410500482529

(2) Bangsbo, J., Mohr, M., & Krustrup, P. Bangsbo J, Mohr M, Krustrup P. (2006).

Physical and metabolic demands of training and match-play in the elite

football player. Journal of Sports Sciences, 24(7), 665-674. https://doi.org/

10.1080/02640410500482529

(3) Chen, Y., Wang, L., Hu, J., & Ye, M. (2020). Vision-Based Fall Event Detection

in Complex Background Using Attention Guided Bi-directional LSTM.

https://doi.org/10.1109/ACCESS.2020.3021795

(4) Gao, M., Li, J., Hong, F., & Long, D. (2019). Day-ahead power forecasting in a

large-scale photovoltaic plant based on weather classification using LSTM.

Energy, 187, 115838.115831-115838.115812. https://doi.org/10.1016/

j.energy.2019.07.168

(5) Li, L., Yang, Y., Yuan, Z., & Chen, Z. (2021). A spatial-temporal approach for

traffic status analysis and prediction based on Bi-LSTM structure. Modern

Physics Letters B, 35(31). https://doi.org/10.1142/S0217984921504819

(6) Petridis, S., Li, Z., & Pantic, M. (2017). End-To-End Visual Speech

Recognition With LSTMs. IEEE. https://doi.org/10.1109/ICASSP.2017.7952625

(7) Li, Y., Ye, F., Liu, Z., Wang, Z., & Mao, Y. (2021). A Short-Term Photovoltaic

Power Generation Forecast Method Based on LSTM. Mathematical Problems

in Engineering. https://doi.org/10.1155/2021/6613123

(8) Shen, Y., Shao, P., Chen, G., Gu, X., & Zhu, J. (2021). An identification

method of anti-electricity theft load based on long and short-term memory

network. Procedia Computer Science, 183(8), 440-447. https://doi.org/10.1016/

j.procs.2021.02.082

(9) Ren, Y., Wang, J., Yang, C., Xiao, C., & Li, S. (2021). Wind and Solar

Integrated Power Prediction Method Research Based on DT-CWT and

LSTM. Journal of Physics Conference Series, 1754(1), 012008. https://doi.org/

10.1088/1742-6596/1754/1/012008

(10) Peng, Y., Liu, X., Wang, W., Zhao, X., & Wei, M. (2019). Image caption model

of double LSTM with scene factors. Image and Vision Computing, 86(JUN.),

38-44. https://doi.org/10.1016/j.imavis.2019.03.003

(11) Qian, F., Chen, L., Li, J., Ding, C., & Wang, J. (2019). Direct Prediction of the

Toxic Gas Diffusion Rule in a Real Environment Based on LSTM.

International Journal of Environmental Research and Public Health, 16(12),

2133. https://doi.org/10.3390/ijerph16122133

(12) Zhou, L., & Bian, X. (2019). Improved text sentiment classification method

based on BiGRU-Attention. Journal of Physics: Conference Series, 1345(3),

032097. https://doi.org/10.1088/1742-6596/1345/3/032097

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

138

5. CONFLICT OF INTEREST

The authors declared that there is no conflict of interest.

REFERENCES

(1) Bangsbo, J., Mohr, M., & Krustrup, P. (2006). Physical and metabolic

demands of training and match-play in the elite football player. J Sports Sci,

24(7), 665-674. https://doi.org/10.1080/02640410500482529

(2) Bangsbo, J., Mohr, M., & Krustrup, P. Bangsbo J, Mohr M, Krustrup P. (2006).

Physical and metabolic demands of training and match-play in the elite

football player. Journal of Sports Sciences, 24(7), 665-674. https://doi.org/

10.1080/02640410500482529

(3) Chen, Y., Wang, L., Hu, J., & Ye, M. (2020). Vision-Based Fall Event Detection

in Complex Background Using Attention Guided Bi-directional LSTM.

https://doi.org/10.1109/ACCESS.2020.3021795

(4) Gao, M., Li, J., Hong, F., & Long, D. (2019). Day-ahead power forecasting in a

large-scale photovoltaic plant based on weather classification using LSTM.

Energy, 187, 115838.115831-115838.115812. https://doi.org/10.1016/

j.energy.2019.07.168

(5) Li, L., Yang, Y., Yuan, Z., & Chen, Z. (2021). A spatial-temporal approach for

traffic status analysis and prediction based on Bi-LSTM structure. Modern

Physics Letters B, 35(31). https://doi.org/10.1142/S0217984921504819

(6) Petridis, S., Li, Z., & Pantic, M. (2017). End-To-End Visual Speech

Recognition With LSTMs. IEEE. https://doi.org/10.1109/ICASSP.2017.7952625

(7) Li, Y., Ye, F., Liu, Z., Wang, Z., & Mao, Y. (2021). A Short-Term Photovoltaic

Power Generation Forecast Method Based on LSTM. Mathematical Problems

in Engineering. https://doi.org/10.1155/2021/6613123

(8) Shen, Y., Shao, P., Chen, G., Gu, X., & Zhu, J. (2021). An identification

method of anti-electricity theft load based on long and short-term memory

network. Procedia Computer Science, 183(8), 440-447. https://doi.org/10.1016/

j.procs.2021.02.082

(9) Ren, Y., Wang, J., Yang, C., Xiao, C., & Li, S. (2021). Wind and Solar

Integrated Power Prediction Method Research Based on DT-CWT and

LSTM. Journal of Physics Conference Series, 1754(1), 012008. https://doi.org/

10.1088/1742-6596/1754/1/012008

(10) Peng, Y., Liu, X., Wang, W., Zhao, X., & Wei, M. (2019). Image caption model

of double LSTM with scene factors. Image and Vision Computing, 86(JUN.),

38-44. https://doi.org/10.1016/j.imavis.2019.03.003

(11) Qian, F., Chen, L., Li, J., Ding, C., & Wang, J. (2019). Direct Prediction of the

Toxic Gas Diffusion Rule in a Real Environment Based on LSTM.

International Journal of Environmental Research and Public Health, 16(12),

2133. https://doi.org/10.3390/ijerph16122133

(12) Zhou, L., & Bian, X. (2019). Improved text sentiment classification method

based on BiGRU-Attention. Journal of Physics: Conference Series, 1345(3),

032097. https://doi.org/10.1088/1742-6596/1345/3/032097

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

(13) Tian, X., & Li, J. (2019). A novel improved fruit fly optimization algorithm for

aerodynamic shape design optimization. Knowledge Based Systems,

179(SEP.1), 77-91. https://doi.org/10.1016/j.knosys.2019.05.005.

(14) Zhang, F. (2019). Research on Improving Prediction Accuracy of Sports

Performance by Using Glowworm Algorithm to Optimize Neural Network.

International Journal of Information and Education Technology, 9(4), 302-305.

https://doi.org/10.18178/ijiet.2019.9.4.1216

(15) Rosenfeld, P. J., Brown, D. M., Heier, J. S., Boyer, D. S., Kaiser, P. K., Chung, C.

Y., Group, M. S. (2004). Ranibizumab for neovascular age-related macular

degeneration. https://doi.org/10.1016/j.ajo.2005.02.003

(16) Krustrup, P., Mohr, M., & Bangsbo, J. (2002). Activity profile and

physiological demands of top-class soccer assistant refereeing in relation

to training status. J Sports, 20(11), 861-871. https://doi.org/

10.1080/026404102320761778

(17) Dooman, C. S., & Jones, D. (2009). Down, But Not Out: In-Season

Resistance Training for the Injured Collegiate Football Player. Strength &

Conditioning Journal, 31(5), 59-68. https://doi.org/10.1519/

ssc.0b013e3181b9983e

(18) A, C. W., A, X. W., Jz, A., Liang, Z. A., Xiao, B. A., Xin, N. B., Ehd, A. (2021).

Uncertainty Estimation for Stereo Matching Based on Evidential Deep

Learning. https://doi.org/10.1016/j.patcog.2021.108498

(19) Zohali, H., Naderi, B., & Mohammadi, M. (2019). The economic lot scheduling

problem in limited-buffer flexible flow shops: Mathematical models and a

discrete fruit fly algorithm. Applied Soft Computing. https://doi.org/10.1016/

j.asoc.2019.03.054

(20) Cai, W., Zhai, B., Liu, Y., Liu, R., & Ning, X. (2021). Quadratic polynomial

guided fuzzy C-means and dual attention mechanism for medical image

segmentation. Displays, 70, 102106. https://doi.org/10.1016/

j.displa.2021.102106

(21) Miao, J., Wang, Z., Ning, X., Xiao, N., Cai, W., & Liu, R. (2022). Practical and

secure multifactor authentication protocol for autonomous vehicles in 5G.

Software: Practice and Experience. https://doi.org/10.1002/SPE.3087

(22) Ning, X., Duan, P., Li, W., & Zhang, S. (2020). Real-time 3D face alignment

using an encoder-decoder network with an efficient deconvolution layer.

IEEE Signal Processing Letters, 27, 1944-1948. https://doi.org/doi.org/10.1109/

LSP.2020.3032277

(23) Ying, L., Nan, Z. Q., Ping, W. F., Kiang, C. T., Pang, L. K., Chang, Z. H., Nam, L.

(2021). Adaptive weights learning in CNN feature fusion for crime scene

investigation image classification. Connection Science. https://doi.org/

10.1080/09540091.2021.1875987

(24) Ning, X., Gong, K., Li, W., & Zhang, L. (2021). JWSAA: joint weak saliency

and attention aware for person re-identification. Neurocomputing, 453,

801-811. https://doi.org/10.1016/j.neucom.2020.05.106

(25) Yan, C., Pang, G., Bai, X., Liu, C., Xin, N., Gu, L., & Zhou, J. (2021). Beyond

triplet loss: person re-identification with fine-grained difference-aware

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

139

pairwise loss. IEEE Transactions on Multimedia. https://doi.org/10.1109/

TMM.2021.3069562

(26) Connor, J. T., Martin, R. D., & Atlas, L. E. (2002). Recurrent neural networks

and robust time series prediction. IEEE Transactions on Neural Networks,

5(2), 240-254. https://doi.org/10.1109/72.279188

(27) Yu, Z., Li, S., Sun, L. N. U., Liu, L., & Haining, W. Multi-distribution noise

quantisation: an extreme compression scheme for transformer according

to parameter distribution. https://doi.org/10.1080/09540091.2021.2024510

(28) Frayssinet, M., Esenarro, D., Juárez, F. F., y Díaz, M. (2021). Methodology

based on the NIST cybersecurity framework as a proposal for

cybersecurity management in government organizations. 3C TIC.

Cuadernos de desarrollo aplicados a las TIC, 10(2), 123-141. https://doi.org/

10.17993/3ctic.2021.102.123-141

(29) Zhang Min, Lu Xuewen, Hoffman Ettiene, Kharabsheh Radwan & Xiao

Qianghua. (2022). Radioactive source search problem and optimisation

model based on meta-heuristic algorithm. Applied Mathematics and

Nonlinear Sciences, 7 (2), 601-630. https://doi.org/10.2478/

AMNS.2021.2.00159.

https://doi.org/10.17993/3ctecno.2023.v12n1e43.127-140

3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143

Ed.43 | Iss.12 | N.1 January - March 2023

140